{"id":"0e709416-e53d-4cdb-a8ef-5fdb3dfc9928","arxiv_id":"2501.09066","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"UMAP pre-selection on seven observed NIRCam colors finds 44 quiescent galaxy candidates at 3<z<6 in JADES with roughly one-fifth to one-half the candidate pool of color-color methods, including young systems those methods miss.","lead":"The authors used a machine-learning method called UMAP to pre-select rare, non-star-forming galaxies from a JWST survey, finding 44 candidates while cutting the number of galaxies needing detailed analysis by about a factor of five compared to standard color cuts. The approach recovers young quiescent galaxies that existing color-color selection misses, which could improve census counts of how galaxies stop forming stars in the early universe.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The pre-selection completeness is anchored to JAGUAR's UVJ-based quiescent locus and a Baker-calibrated radius; young blue quiescent galaxies outside that locus are untested, so the number densities rest on an unquantified selection function.","rationale":"The central methodological claim, that UMAP pre-selection is more efficient than color-color wedges and recovers quiescent galaxies including young ones, is supported by the figures and the comparison with Long et al. (2024). The Baker et al. validation shows that the method works for the massive, spectroscopically confirmed population. However, the load-bearing assumption is that the JAGUAR training set spans the color space of real high-redshift quiescent galaxies, including young, blue, recently quenched systems. The paper explicitly concedes that JAGUAR does not cover this diversity (Section 3.2), and the selection radius is calibrated on a sample that also does not cover it. Consequently, the efficiency numbers are conditional on training-set coverage, and the number densities, which are presented without completeness corrections, inherit this condition. The reader's weakest assumption identifies the same issue; my concern reinforces it by pointing to the specific in-sample calibration of the radius and the absence of any recovery test for the young blue population. This does not overturn the main efficiency result for the locus JAGUAR represents, but it means the sample should be treated as incomplete for abundance claims until tested. The paper is a solid new application and the CONDITIONAL verdict is appropriate.","tokens_in":27258,"tokens_out":4958,"duration_ms":55531,"concrete_test":"Retrain the published UMAP pipeline on an augmented mock sample that adds quiescent SEDs from an independent library (e.g., beagle or CIGALE models with sSFR < 0.2/t_U, log M* = 8-10, ages 0.05-0.3 Gyr, and E(B-V) < 0.3) to the JAGUAR training set, then inject 500 such galaxies with JADES noise and photometric scatter into the 43,926-galaxy parent sample. Run the radial-distance pre-selection of Section 5.1 unchanged and measure the recovery fraction of the injected young quiescent population inside the radius-1 pool. If recovery is below about 80%, or if it concentrates only in the old locus, the published pool and number densities are incomplete for exactly the population the paper claims to add. Also recompute Table 2 applying a completeness correction based on this recovery fraction and check whether the z>4 number densities shift by more than the quoted uncertainties.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 3.2 acknowledges that JAGUAR's quiescent SEDs are drawn from a UVJ-selected, mostly massive parent sample and do not cover the full diversity of z>3 quiescent populations. The UMAP atlas and the radius-1 thresholds in Section 5.1 are therefore built around the color locus JAGUAR already knows. The radius is explicitly calibrated by doubling the 0.5 radius that encloses the 17 Baker et al. (2024) galaxies, which are massive (M* > 10^10 Msun) and spectroscopically confirmed, precisely the population JAGUAR represents. The subsequent recovery of 44 candidates, including some with mass-weighted ages below 300 Myr, does not demonstrate that similarly young blue systems outside this locus are captured; it only shows that some young-ish candidates happen to share the old-quiescent locus. The candidate pool of 2,282 and the claim of about 5x efficiency are conditional on this training-set coverage. Because the number densities in Table 2 and Figure 7 are raw counts over the selected pool with no completeness correction (Section 5.6), any missed young blue quiescent galaxies would directly bias the high-redshift and low-mass abundances downward. The paper even notes that the exact pre-selection method is arbitrary (Section 5), so the efficiency numbers are not uniquely defined.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper applies the Uniform Manifold Approximation and Projection (UMAP) technique to seven observed-frame NIRCam colors of 43,926 galaxies in JADES (GOODS-N and GOODS-S) to pre-select quiescent galaxy candidates at 3<z<6. A UMAP embedding is trained on ~75,000 JAGUAR mock galaxies, and the observed sample is mapped onto this embedding. Two pre-selection strategies are presented: a radial cut of radius 1 around the z>=3 quiescent model cluster, yielding 2,282 galaxies and 44 quiescent candidates, and two rectangular cuts enclosing the main model clusters, yielding 247 galaxies and 29 candidates. The paper claims that the UMAP-based pre-selection is roughly five times more efficient than an equivalent observed-frame color-color wedge and about twice as efficient as the wide Long et al. (2024) wedge, and it reports number densities at 3<z<6 that agree with the literature at z<4 and are generally higher at z>4. The final catalog contains 27 newly discovered candidates, several with mass-weighted ages below 300 Myr.","tokens_in":27542,"tokens_out":4083,"duration_ms":44009,"significance":"If the efficiency claims hold, the paper provides a genuinely useful tool for finding rare high-redshift quiescent galaxies in JWST surveys, and the public code and catalog are a valuable community resource. The efficiency comparison in Figure 6 is well constructed and clearly demonstrates that, for the specific candidate set found here, UMAP pre-selection yields a smaller pool than a color-color wedge containing the same candidates. The recovery of 17 spectroscopically confirmed Baker et al. (2024) quiescent galaxies near the model cluster is encouraging, and the new low-mass, young candidates are of interest. The main weakness is that the pre-selection completeness is anchored to JAGUAR's UVJ-based quiescent locus and to the Baker et al. sample used for calibration, so the selection function is not independently established. This limits the interpretation of the number densities, which are presented without a completeness correction.","major_comments":[{"comment":"The number densities in Table 2 and Figure 7 are based on raw counts over the pre-selected pool with no completeness correction, but the pre-selection is trained on JAGUAR quiescent SEDs that, as the paper states in Section 3.2, are drawn from a UVJ-selected, mostly massive parent sample and do not cover the full diversity of z>3 quiescent galaxies, especially bluer, younger populations. If such populations lie outside the model-based clusters in UMAP space, the pre-selection will miss them, and the z>4 and low-mass number densities will be biased low. The statement that the impact is negligible is not demonstrated. Please provide a quantitative completeness test, for example by injecting young, blue quiescent SEDs (from other models or from the observed candidates themselves) into the UMAP embedding and measuring the recovery fraction, or explicitly present the number densities as lower limits with the selection function stated as unknown.","section":"Section 3.2 and Section 5.6"},{"comment":"The validation using the 17 Baker et al. (2024) spectroscopically confirmed galaxies is partly circular. The UMAP hyperparameters (n_neighbors=100, min_dist=0.01) were chosen by sweeping to maximize clustering of the z>=3 quiescent JAGUAR models, and the radial cut of 1 was chosen by doubling the radius of 0.5 that encloses those same 17 galaxies. Consequently, the fact that all 17 lie close to the quiescent model cluster is not an independent confirmation of the technique's completeness. An out-of-sample test, such as applying the pre-trained UMAP to a survey not used in the calibration (e.g., CEERS or UNCOVER) or using a held-out subset of spectroscopically confirmed galaxies, would strengthen the claim that the method recovers quiescent galaxies in general rather than just the specific population used for tuning.","section":"Sections 4.2, 4.3.1, and 5.1"},{"comment":"The headline efficiency comparison ('about five times fewer galaxies') is constructed retrospectively: the color-color wedge that encloses all 44 candidates is obtained by adjusting the intercepts of the Long et al. (2024) criteria, and the rectangular UMAP cuts are also defined a posteriori to enclose the model clusters. The paper itself notes in Section 5 that the exact pre-selection method is arbitrary. Please add a sensitivity analysis showing how the candidate-pool sizes and the efficiency ratios change with reasonable variations of the UMAP hyperparameters and the radial/rectangular cut positions, so that the reader can assess whether the factor-of-five and factor-of-two efficiency gains are robust or are a product of the specific choices made here.","section":"Section 5.5"}],"minor_comments":[{"comment":"In the paragraph on photometric redshift accuracy, 'redshfits' should be 'redshifts'.","section":"Section 2.2"},{"comment":"The rectangular cut for Region C is given as -4.0 <= UMAP1 <= -2.8 and -4.8 <= UMAP2 <= -4.6, which is visibly wider than the small group of five models at approximately (-3.5, -5); please state explicitly how the rectangle boundaries were chosen and whether the result is sensitive to those boundaries.","section":"Section 5.2"},{"comment":"The caption states that selecting all 44 candidates using color-color cuts would require the extended wedge shown in magenta, but it should be clearer that this is a custom wedge constructed for this comparison and not the Long et al. (2024) selection, which is shown by the other lines.","section":"Figure 6 caption"},{"comment":"Several entries have extremely low log sSFR values with very large asymmetric uncertainties (e.g., ID 13124 and ID 40382); consider flagging these as poorly constrained or excluding them from the number-density computation to avoid giving them equal weight.","section":"Table 1"},{"comment":"The number-density computation perturbs photometric redshifts using the 16th and 84th percentiles of the PDF, but the selection criteria in Section 5.1 also require the 16th percentile to satisfy z>=2.5; please clarify whether the Monte Carlo realizations maintain consistency with the sample selection criteria.","section":"Section 5.6"}],"recommendation":"major_revision","confidential_remarks":"The paper is suitable in scope for ApJ and the central methodological comparison is sound, but the lack of an independent completeness test is a genuine load-bearing issue for the number-density claims. A revision that adds a recovery simulation for young, blue quiescent galaxies and a sensitivity analysis of the UMAP parameters would resolve the main concern. I would not reject the paper on the circularity issue alone, since the efficiency comparison with color-color wedges is a valid retrospective exercise, and the candidate catalog has independent value."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a solid methodological contribution, not a breakthrough. The UMAP-based pre-selection works, is new at z>3, and the efficiency gain over color-color wedges looks real. The main caveat is that the validation is partly circular and the completeness of the selection is not fully quantified, so treat the number densities as suggestive.\n\nWhat's new: applying UMAP to seven observed NIRCam colors to pre-select quiescent galaxies over 3<z<6. The comparison with Long et al. (2024) is well done, and the result that you can get 44 candidates from ~2,300 galaxies (vs ~10,500 with an extended color wedge) is a real practical gain. The recovery of young (<300 Myr) quiescent candidates that color-color misses is interesting. The paper is transparent about the arbitrary choice of pre-selection method and releases code and the catalog, which is good.\n\nSoft spots: The main one is the circularity in the validation. The radius of 1 in UMAP space is set as twice the radius that encloses the Baker et al. spectroscopically confirmed galaxies. Those are massive, old quiescent galaxies, and they are also the population that JAGUAR knows well. So the validation shows that the method recovers the type of object it was tuned to find. It doesn't test whether young, blue, recently quenched galaxies outside the JAGUAR locus are captured. The authors acknowledge this limitation of JAGUAR but then brush it off with 'our success shows the impact is negligible' – that's not a rigorous argument. The number densities in Section 5.6 are raw counts with no completeness correction, so they could be biased low if some young blue quiescent galaxies are missed. The authors note this indirectly by saying the measurements are consistent with literature, but it should be stated more clearly.\n\nAnother minor point: the efficiency comparison in Figure 6 is fair, but it depends on the specific color-color wedges chosen. The authors show an extended wedge to capture all 44, but one could imagine other color cuts. The claim is still reasonable, though.\n\nOverall: the central efficiency result holds up. The sample is useful, the method is well described, and the code is provided. It should be reviewed. The authors should address the circularity and completeness concerns before publication, but they don't invalidate the main point.\n\nRecommendation: send to a serious referee. The paper is worth engaging with.","headline":"A useful and mostly solid application of UMAP to pre-select z>3 quiescent galaxies; the efficiency gain is real, but the validation is partly circular and the number densities need completeness caveats.","tokens_in":28141,"tokens_out":2175,"would_cite":true,"duration_ms":21070,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Manifold learning finds 44 quiescent galaxies at z>3 using a candidate pool five times smaller than color-color selection.","keywords":["Quenched galaxies","High-redshift galaxies","Astronomy data visualization","quiescent galaxies","UMAP","manifold learning","JWST/NIRCam","JADES"],"falsifier":"Take a spectroscopically complete sample of galaxies at $3 < z < 6$ in JADES or a similar field, measure their sSFRs, and compare the quiescent ones to the UMAP candidate pools; if any spectroscopically confirmed quiescent galaxy falls outside both the radial-distance and rectangular pre-selection regions, the method is incomplete as a pre-selection and the reported number densities would be lower limits rather than measurements.","tokens_in":27060,"feed_emoji":"🔭","tokens_out":16939,"duration_ms":125787,"temperature":0.7,"pith_summary":"Quiescent galaxies at $3 < z < 6$ are rare and hard to find: they formed and stopped forming stars within about two billion years of the Big Bang, and existing rest-frame color-color diagrams become unreliable at these redshifts. This paper argues that a machine-learning pre-selection step can replace most of that search. The authors train the Uniform Manifold Approximation and Projection algorithm (UMAP) on seven observed-frame NIRCam colors from the JAGUAR mock catalog, producing a two-dimensional map on which galaxies with similar colors sit close together. Projecting 43,926 JADES galaxies onto this map and fitting only those near the quiescent model clusters yields 44 quiescent candidates from a pool of about 2,300 galaxies, roughly five times smaller than the pool needed with observed-frame color-color wedges. Two-thirds of the candidates are recovered from a pool of 247 galaxies, and nearly two-thirds are newly discovered, including galaxies with mass-weighted ages below 300 Myr.","feed_headline":"Machine learning finds 44 z>3 quiescent galaxies with 5x smaller pool","feed_subtitle":"UMAP on seven NIRCam colors cuts the pre-selection pool from ~10,500 to ~2,300 and catches young <300 Myr systems.","key_machinery":"The central object is UMAP (Uniform Manifold Approximation and Projection), a nonlinear dimensionality-reduction algorithm that builds a weighted nearest-neighbor graph in a high-dimensional space and then embeds it into two dimensions while preserving local and global structure. Trained here on roughly 75,000 JAGUAR mock galaxies described by seven observed-frame NIRCam colors (F115W-F150W, F115W-F277W, F150W-F200W, F150W-F277W, F200W-F277W, F200W-F356W, F277W-F444W), the embedding places galaxies with similar colors near one another. The map is used as a lookup table: observed JADES galaxies are projected onto it with UMAP's transform routine, and proximity to the clustered quiescent models defines the candidate pool. The quiescence of the pre-selected galaxies is then established by fitting their HST+JWST photometry with the bagpipes SED-fitting code and applying an sSFR threshold of 0.2 divided by the age of the Universe at that redshift.","core_discovery":"The paper's central claim is that a pre-selection based on UMAP manifold learning is a more efficient way to find quiescent galaxies over $3 < z < 6$ than observed-frame color-color diagrams. Using seven NIRCam colors designed to isolate high-redshift quiescent galaxies, the trained UMAP transformation maps the 62 model quiescent galaxies at $z \\geq 3$ into a tight cluster; all 17 spectroscopically confirmed massive quiescent galaxies from a recent JWST study fall in the same region, validating the map. When the full JADES observational sample is transformed, selecting all galaxies within a radial distance of 1 in UMAP space yields 2,282 candidates, of which 44 are quiescent after SED fitting. An even smaller selection using two rectangular regions recovers 29 of the 44 from only 247 galaxies, about twice as efficient as the comparable color-color wedge. The paper concludes that the method captures young ($< 300$ Myr) quiescent galaxies that color-color criteria tend to miss, and that the derived number densities agree with earlier work at $z < 4$ while tending to be higher, but consistent within errors, at $z > 4$.","pith_inferences":["Adding more photometric bands or morphological information as UMAP inputs could shrink the candidate pool further, a direction the paper does not explore.","Because JAGUAR's quiescent templates are drawn from UVJ-selected galaxies, the map may be blind to the bluest, youngest quiescent galaxies; injecting such synthetic SEDs into the training set would test whether they land inside the candidate regions.","The method's reliance on observed-frame colors means it can be applied without photometric redshifts, so it could be run on raw photometric catalogs before any redshift estimation is performed.","If applied to wide-area surveys, the pre-selection would drastically reduce the computational cost of SED fitting, potentially enabling quiescent galaxy searches over hundreds of square degrees."],"forward_implications":["The same trained UMAP map can be applied to other JWST surveys with NIRCam coverage, so future searches for $z > 3$ quiescent galaxies can avoid SED-fitting tens of thousands of sources.","If the higher number densities at $z > 4$ hold, they would strengthen the already serious tension between observed quiescent galaxy abundances and cosmological simulations, which currently underpredict them by 1-2 dex.","The recovery of several candidates with mass-weighted ages below 300 Myr implies that a subset of galaxies quench within about 300 Myr of their formation, providing a sharp timescale constraint on quenching mechanisms at early epochs.","The roughly fivefold reduction in the candidate pool makes it feasible to obtain spectroscopy for the vast majority of pre-selected candidates, which is needed to confirm their quiescence and measure their ages."],"supporting_citations":[{"why":"Provides the JAGUAR mock catalog whose mock photometry trains the UMAP algorithm and defines the quiescent model sample.","marker":"Williams et al. 2018"},{"why":"Describes the UMAP algorithm used to build the two-dimensional map and to transform observed colors.","marker":"McInnes et al. 2018"},{"why":"Supplies the observed-frame color-color wedges that serve as the baseline for the efficiency comparison.","marker":"Long et al. 2024"},{"why":"Supplies spectroscopically confirmed massive quiescent galaxies at z ≥ 3 used to validate the UMAP map's placement of real quiescent galaxies.","marker":"Baker et al. 2024"},{"why":"Provides the bagpipes SED-fitting tool used to measure stellar masses and sSFRs for the pre-selected candidates.","marker":"Carnall et al. 2018"},{"why":"Provides the JADES NIRCam photometry and source catalogs from which the observational sample is drawn.","marker":"Rieke et al. 2023"}],"fun_headline_variants":["UMAP pre-selects 44 quiescent galaxies from 2,300 in JWST data","Manifold learning finds 44 quiescent galaxies with 5x smaller pool","New ML method spots young quiescent galaxies color-color criteria miss","JWST+UMAP: 44 rare quiescent galaxies found in a tiny pool","UMAP cuts quiescent galaxy search pool to 2,300, finds 44"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the JAGUAR mock catalog contains the full range of observed colors of real high-redshift quiescent galaxies, particularly the young, blue, recently quenched systems; the authors acknowledge that JAGUAR's quiescent SEDs are built from UVJ-selected, mostly massive galaxies and may not cover the bluest population, and if that diversity is missing, the UMAP map will place such galaxies away from the quiescent model clusters and the pre-selection will miss them.","fun_headline_variants_meta":{"raw":{"variants":["UMAP pre-selects 44 quiescent galaxies from 2,300 in JWST data","Manifold learning finds 44 quiescent galaxies with 5x smaller pool","New ML method spots young quiescent galaxies color-color criteria miss","JWST+UMAP: 44 rare quiescent galaxies found in a tiny pool","UMAP cuts quiescent galaxy search pool to 2,300, finds 44"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000297,"raw_usage":{"total_tokens":1810,"prompt_tokens":1124,"completion_tokens":686,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":740,"completion_tokens_details":{"reasoning_tokens":574}},"tokens_in":740,"tokens_out":686,"duration_ms":7463,"temperature":1.0,"reasoning_tokens":574,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T20:10:18.121471+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a spectroscopically complete sample of galaxies at $3 < z < 6$ in JADES or a similar field, measure their sSFRs, and compare the quiescent ones to the UMAP candidate pools; if any spectroscopically confirmed quiescent galaxy falls outside both the radial-distance and rectangular pre-selection regions, the method is incomplete as a pre-selection and the reported number densities would be lower limits rather than measurements.","supporting_citations":[],"review_version":1}